Automated mammography report generation using vision-language models achieves high metric scores through increased model capacity rather than synthetic pretraining. The research investigates a two-stage synthetic-to-real training pipeline designed to overcome a shortage of publicly available paired medical images and free-text radiology reports.
Data Scarcity in Mammography Training
Publicly available mammography collections such as VinDr-Mammo and CBIS-DDSM provide structured annotations like Breast Imaging Reporting and Data System (BI-RADS) categories, but they lack the free-text radiology reports required to supervise vision-language models. To address this data scarcity, researchers developed a label-to-report synthesis pipeline that converts structured BI-RADS annotations into clinical text, according to researchsquare.com. The synthesis stage generated 400 reports and retained 390 to pretrain the MedGemma-4B model using multimodal Low-Rank Adaptation (LoRA), before fine-tuning it on 407 real radiologist reports from the Digital Database for Screening Mammography (DMID) corpus.
Model Outperforms Nine External Baseline Methods
Evaluated on a leak-free DMID test subset of 49 samples, the model reaches a ROUGE-L score of 0.656, METEOR of 0.666, CIDEr of 0.649, and BERTScore of 0.917, exceeding every external baseline. Complementary data notes that a test set of 52 samples achieved a BLEU-4 of 0.312, ROUGE-L of 0.671, METEOR of 0.685, CIDEr of 0.729, and BERTScore of 0.917. This performance outperforms nine baseline methods, including zero-shelf vision-language models like Qwen2.5-VL-7B, Qwen3-VL-8B, and LLaVA-1.6-7B, as well as pipeline approaches such as CLIP+GPT2 and MedCLIP+GPT2.
Model Capacity Versus Synthetic Pretraining
Further controlled evaluations across 60 training runs demonstrate that synthetic pretraining does not drive performance gains. The apparent advantage initially measured between models stemmed from unequal adapter capacity. When matched for capacity, both two-stage training and direct fine-tuning yield equivalent results. Across three corpus sizes, three capacities, and five initializations against a predetermined margin, no configuration favors two-stage training.
Diagnostic Accuracy Challenges in Clinical Reporting
Evaluation reveals a divergence between textual fluency and clinical quality within the models. Under an assertion-level criterion, no model fabricates findings on normal studies, with failure modes centering entirely on omission. However, where two-stage training gains 0.058 in ROUGE-L text matching, BI-RADS agreement and sensitivity both decrease. With agreement never exceeding $\kappa = 0.14$, the primary obstacle to clinically useful reporting remains diagnostic accuracy rather than linguistic fluency.
Keep reading